Latent Visual Reasoning
Overview
Latent visual reasoning explores whether models can preserve and manipulate visual evidence in compact internal states instead of spelling out every step in text or repeatedly processing images. Common approaches add continuous visual tokens or latent workspaces, train them with visual-region supervision, answer-based rewards, or distillation, and use predictive world models to anticipate future scenes. These methods are tested on fine-grained perception and spatial benchmarks such as V* and BLINK, as well as video understanding and embodied planning tasks, often aiming to reduce inference time and token use. A central open question is whether the latent states causally contribute grounded evidence to answers, since accuracy gains can also arise from formatting, attention changes, or auxiliary training; interventions and evidence-grounding remain important evaluation tools.
Latest papers
90 papers11 papers in the last four weeks, up 120% on the four weeks before. 0.1% of all new papers.